Noisy speech recognition is an involved and complicated research area that lacks both sound theory and experiences. Early experiments showed that word recognition performance dropped significantly to noise ratios (SNR) less than 24dB (1) and accuracy reduced by 34% for a SNR of 11 dB (2). Although a number of speech enhancement algorithms have been proposed for noise reduction (2,3), none of them seems to be an effective solution, especially when the SNR is relatively low.
Bibliographic reference. Chen, Xixian / Cai, Changnian (1989): "A novel approach for degraded speech recognition", In EUROSPEECH-1989, 1594-1595.